Neural Networks and Deep Learning

Lecture 10 · Game: You or the agent? A battery on the Romanian day-ahead market

You or the agent? A battery on the Romanian day-ahead market

You run a 2 MWh battery on the Romanian day-ahead electricity market. Each card is a real day of 2025 or 2026. Like the agent of the lecture, you plan the day before its prices are known: you see only yesterday's 24 hourly prices and choose one of four simple plans. Then you see what your plan, the agent and a trader with perfect foresight earned on that day's real prices.

The battery charges or discharges 1 MWh of stored energy per hour, starts the day empty and loses 15% of the energy over a full cycle (85% round-trip efficiency). The agent is the tabular Q-learning agent of the lecture (reinforcement learning: it learned which hours to charge and discharge from 1,746 days of 2019-2023); its state is the hour, the stored energy and the rank of that hour among yesterday's prices. A card is won against the agent when your plan earns at least as much as the agent; the score counts the cards on which you chose the best of the four plans.

Class board for the projector

Data: hourly day-ahead prices for Romania, energy-charts.info (Fraunhofer Institute for Solar Energy Systems, ISE), up to 31 October 2026; hours in Central European Time (CET). The 40 days of the game are drawn at random from the 638 test days (seed 20261108); every game draws 8 of them. The hours of the fixed plans come from the mean price profile of 2019-2024. Code: lectures/L10/make_game_L10.py.